Structural damage diagnosis method based on optimized stacked autoencoder and multi-signal fusion
The method addresses hyperparameter and sensor data inconsistencies in structural damage diagnosis by using optimized stacked autoencoders and multi-signal fusion, enhancing accuracy and reliability through improved whale optimization.
Patent Information
- Application Number
- CN202111387865.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-22
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2041-11-22
AI Technical Summary
Existing deep learning models for structural damage diagnosis are limited by hyperparameter settings and sensor placement uncertainties, and existing technologies fail to accurately address the challenges of hyperparameter configuration and sensor data inconsistencies, leading to inaccurate and unreliable damage diagnosis results.
A method utilizing optimized stacked autoencoders and multi-signal fusion, enhanced by an improved whale optimization algorithm, to optimize hyperparameters and integrate sensor data for improved accuracy and reliability.
The method enhances the accuracy and reliability of structural damage diagnosis by optimizing hyperparameters and fusing sensor data, achieving higher recognition rates and improved structural health assessment.
Smart Images

Figure CN114118249B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of damage diagnosis in structural health monitoring, and in particular to a structural damage diagnosis method based on an optimized stacked autoencoder deep learning model and multi-sensor measurement signal fusion. Background Art
[0002] Health monitoring of civil engineering structures such as buildings and bridges is an important means to ensure the safe operation and scientific maintenance of the corresponding infrastructure. The health monitoring method based on structural vibration response has been widely studied by scholars at home and abroad due to its convenience and feasibility, and has been applied to a certain extent in engineering. Changes in physical properties such as reduced stiffness caused by structural damage will lead to changes in structural vibration characteristics. Therefore, data mining and analysis of structural vibration responses measured under different structural states can achieve structural health assessment and damage diagnosis. Structural health monitoring methods mainly include two categories: physical model-based and data-driven. Among them, the physical model-based method requires the establishment of an accurate structural simulation model, which is often difficult to implement in practice. The data-driven method uses the measured structural time domain and frequency domain responses, combined with effective data processing methods, to construct a mapping between vibration response characteristics and structural health status, thereby achieving the purpose of damage diagnosis. With the development of computer science, research on structural health monitoring based on massive vibration monitoring data and machine learning methods has received great attention.
[0003] As an important subfield of machine learning, deep learning has achieved remarkable results in the fields of structural health monitoring in recent years. The deep learning methods used in the field of structural damage diagnosis mainly include recursive neural network (RNN), generative adversarial network (GAN), regional convolutional neural network (R-CNN) and stacked autoencoder (SAE). Among them, SAE is one of the most commonly used methods for extracting damage features in structural vibration signal processing.
[0004] Although the above first-order deep learning algorithms show good performance in structural damage diagnosis, their recognition accuracy is affected by many factors. Among them, the model hyperparameter setting involved in the deep learning algorithm is the main influencing factor. Deep learning models established according to different hyperparameters will lead to very different recognition results. Therefore, how to configure the optimal values of the algorithm parameters is a major difficulty in the application of deep learning methods in structural damage diagnosis. The whale optimization method (WOA) is a novel heuristic optimization algorithm proposed by Mirjalili and Lewis. It is based on the mathematical simulation of the hunting behavior of humpback whales, and the position of each humpback whale represents a feasible solution. The mathematical model of WOA mainly includes three behaviors: surrounding prey, bubble net utilization, and prey search. Although WOA is widely used to solve various engineering problems due to its easy implementation, it has the problems of slow convergence and low computational accuracy when dealing with complex optimization problems.
[0005] In addition, in order to achieve an accurate and comprehensive assessment of the structural status, it is necessary to collect and analyze vibration signals from multiple sensors at different locations. However, due to the influence of environmental uncertainties such as sensor failure or deployment location, the diagnostic results of different sensors may be different, and sometimes even contradictory, which adds challenges to the accurate assessment of the structural health status. Summary of the invention
[0006] In order to avoid the problem of inaccurate damage diagnosis results caused by the hyperparameter settings of the above-mentioned deep learning model and the differences in the analysis results of the measurement signals of each sensor, the present invention proposes a structural damage diagnosis method based on optimized stacked autoencoders and multi-signal fusion, thereby improving the accuracy of structural damage diagnosis.
[0007] The present invention is a structural damage diagnosis method based on optimized stacked autoencoders and multi-signal fusion, characterized in that the specific steps are as follows:
[0008] S1: Measure and collect vibration signals of the structure under different health states. Accelerometers are deployed at multiple monitoring points on the structure to measure the vibration acceleration data of the structure at a specific position and direction, and the data of n sensor measurement channels corresponding to Ncc different structural health states are obtained. The number of measurements in each state is M.
[0009] S2: Preprocess the acquired monitoring response data, mainly using Fourier transform to obtain frequency domain response, and use principal component analysis to denoise and reduce the dimension of the frequency domain response data. By discarding the principal components with small contributions, the redundant information in the original data set is eliminated, and the principal component data set of the frequency domain response is obtained.
[0010] S3: For the principal component data set of the signal of each sensor in S2, a stacked autoencoder (EWOA-DSAE) deep learning model is constructed for feature analysis by using the enhanced whale optimization algorithm (EWOA) to optimize the network model parameters. The stacked autoencoder model includes an encoder and a decoder, which are respectively composed of input / output layers, hidden layers, and mapping functions between them. The hyperparameters of the model include the number of neurons in each hidden layer HLN, the weight penalty factor λ, the dropout probability dr, and the learning rate lr, which are optimized and determined by the proposed enhanced whale optimization algorithm. The deep learning model EWOA-DSAE-i (i=1,2,…,n) corresponding to each sensor data is constructed using the optimized hyperparameters to calculate the probability value of accurate prediction and the preliminary diagnosis results of structural damage.
[0011] S4: The predicted probability corresponding to the measurement signal of each sensor obtained in S3 is used as the basic probability assignment in the Dempster-Shafer (DS) fusion algorithm to form the final decision on the structural damage diagnosis analysis.
[0012] The enhanced whale optimization algorithm (EWOA) described in step S3 is an improvement on the traditional whale optimization algorithm, solving the problems of slow convergence speed and low calculation accuracy in the traditional method. The enhanced method has been improved in two aspects, including:
[0013] 1) An adaptive inertia weight is added to the original whale optimization algorithm to enhance the local search capability of the algorithm. The adaptive inertia weight can be expressed as formula (1).
[0014]
[0015] Among them, iw max and iw min Represent the maximum and minimum inertia weights respectively; t represents the current iteration number; T max represents the maximum number of iterations. From formula (1), it can be seen that the inertia weight decreases nonlinearly with the increase of the number of iterations. In the initial iteration stage, the inertia weight value is large, so the algorithm has a good global search ability, which enables the whale to quickly approach the approximate area of the prey (the best solution); then, in the later stage of the iteration, the value of the inertia weight gradually decreases, indicating that it has a good local search ability and can effectively improve the accuracy of the optimization results.
[0016] 2) The Archimedean spiral curve is used to replace the logarithmic spiral curve in the original method to increase its search space; the mathematical expression of the Archimedean spiral curve is formula (2):
[0017]
[0018] Among them, a0, b are constants that define the spiral shape; l represents a random number between 0 and 1;
[0019] The new whale bubble network utilization behavior can be expressed as formula (3):
[0020]
[0021] Among them, X*(t) represents the best position of the whale at present, that is, the optimal result; X(t) represents the current position of the whale; a is a variable whose value decreases from 2 to 0; r represents a random number between 0 and 1; p is a random number between 0 and 1; A is a coefficient vector. .
[0022] In step S3, the expression of whale prey search is as follows:
[0023] X(t+1)=X rand -A·|2r·X rand -X(t)| (4)
[0024] Among them, X rand represents the position of a randomly selected whale; if |A|<1, the whale will perform bubble net utilization behavior; if |A|≥1, the whale will perform prey search behavior.
[0025] Based on the above analysis, the calculation process of the enhanced whale optimization algorithm includes:
[0026] S3.1 Setting the size of the whale group N p and the maximum number of iterations N max , initialize the algorithm parameters: a, r, l, a0, b and p;
[0027] S3.2 Calculate the fitness values of Np whales and compare them to get the best individual with the best fitness, and use X * express;
[0028] S3.3 Enter the main loop of the algorithm. If |A| < 1, the position of the whale is updated according to formula (3); if |A| ≥ 1, the position of the whale is updated according to formula (4);
[0029] S3.4 Re-evaluate the fitness values of all whales in the group to find the global optimal individual with the corresponding position. If the current optimal fitness is better than the previous one, replace the previous optimal position with the current optimal position. Otherwise, keep the optimal individual and position unchanged;
[0030] S3.5 checks the algorithm stopping criteria, using the maximum number of iterations as the termination criteria. If the current number of iterations is greater than the maximum number, the algorithm terminates; otherwise, go to step S3.2 and continue the algorithm iteration;
[0031] S3.6 outputs the optimal solution X * .
[0032] The optimization of the DSAE model hyperparameters in step S3 requires setting a reasonable objective function. This method uses the model loss function to design the objective function. The objective function can be expressed as formula (5).
[0033]
[0034] in and k Represents the prediction and true situation of the kth sample; x k represents the kth input; N tr represents the number of training samples; I{·} is the indicator function, which takes a value of 1 when the input is true, and a value of 0 otherwise. It represents the probability that the prediction result based on softmax classification belongs to the true label, and its expression is formula (6):
[0035]
[0036] in, N represents the connection weight and bias vector of the i-th neuron in the output layer corresponding to the i-th structural health state condition; cc Represents the total number of structural health conditions.
[0037] In step S4, the probability results of the EWOA-DSAE-i model corresponding to each sensor signal obtained in S3 are used as separate evidence for constructing basic probability assignment in the DS algorithm. The combination of evidence of multiple sensors deployed at different locations can effectively improve the diagnostic accuracy of the structural state.
[0038] The beneficial effects of the present invention are as follows:
[0039] 1) An enhanced whale optimization algorithm is proposed to optimize the hyperparameters required for the stacked autoencoder model and is applied to structural damage diagnosis. Compared with the traditional whale optimization algorithm and particle swarm optimization algorithm, this optimization method can determine more appropriate hyperparameters.
[0040] 2) Compared with machine learning models such as one-dimensional convolutional neural networks and support vector machines with fixed hyperparameters, the proposed enhanced whale optimized-stacked autoencoder (EWOA-DSAE) has better feature extraction capabilities and higher structural damage diagnosis accuracy.
[0041] 3) Enhanced whale optimization-the combination of stacked autoencoders and DS-based multi-signal fusion algorithm can improve the accuracy of diagnosis and enhance the reliability of structural damage diagnosis even when the number of sensor deployments is limited and the uncertainty is high. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 This is the test model diagram of the cantilever beam structure;
[0043] Figure 2 Diagram of sensor arrangement on the structure;
[0044] Figure 3 This is a schematic diagram of the vibration test process;
[0045] Figure 4 3D t-SNE scatter plot of the original input features;
[0046] Figure 5 It is the optimization process diagram of the number of neurons in the first hidden layer HLN1;
[0047] Figure 6 It is the optimization process diagram of the number of neurons in the second hidden layer HLN2;
[0048] Figure 7 It is the optimization process diagram of the number of neurons in the third hidden layer HLN3;
[0049] Figure 8 Optimization process diagram for discarding probability dr;
[0050] Fig. 9 It is the optimization process diagram of the weight penalty factor λ;
[0051] Fig.10 It is the optimization process diagram of the learning rate lr;
[0052] Fig.11 The convergence of the three optimization methods in the hyperparameter optimization process;
[0053] Fig.12 3D t-SNE scatter plot of the output features of the third hidden layer of the EWOA-DSAE model;
[0054] Fig.13 Confusion matrix diagram for model validation after considering 3-sensor fusion;
[0055] Fig.14 Confusion matrix diagram for model validation after considering 4-sensor fusion;
[0056] Fig.15 This is an implementation flow chart of the structural damage diagnosis method based on optimized stacked autoencoders and multi-signal fusion described in the present invention. DETAILED DESCRIPTION
[0057] The implementation of the structural damage diagnosis method of the present invention is further described below in conjunction with the accompanying drawings.
[0058] Embodiment: In the first step, structural vibration acceleration response data is obtained through sensors and a health monitoring system.
[0059] Attached Figure 1 The figure shows a test model made based on a beam unit component of a real bridge. The test piece is a steel-concrete composite cantilever beam with a total length of 1600mm. There is a vertical support point in the middle of the beam 1000mm away from the cantilever end. Figure 2 Therefore, 10 piezoelectric accelerometers (A1 to A10) are arranged at the following positions. Figure 2 Cracks are applied at the crack positions shown to construct different degrees of structural states. According to the applied crack conditions, a total of 5 structural states are set, namely (1) damage-free state CS1: no cracks; (2) small damage state CS2: crack length 75mm; (3) medium damage state CS3: crack length 150mm; (4) large damage state CS4: crack length 225mm; and (5) severe damage state CS5: crack length 270mm. In each health state, according to the attached Figure 3 The vibration test was carried out according to the test process to collect vibration data of the structure at different positions. 150 hammer tests were carried out at a fixed hammer point, and the experimental data were used for model training; 10 hammer tests were carried out at the other 5 different hammer points, and the measured data were used as verification data for model verification. The length of the time domain data measured in each test was 16,000, and finally 1,000 vibration data collection tests were carried out, and the ratio of training samples to verification samples was 3:1.
[0060] In the second step, the measured time domain response data is converted into frequency domain response. The data of each frequency domain response group is 4000. Using principal component analysis, the frequency domain response is projected into principal components with a data length of 200. The frequency domain response principal components will be used as the input of the stacked autoencoder model. Corresponding to the five structural health states, the output variables of the model are defined as (1,0,0,0,0), (0,1,0,0,0), (0,0,1,0,0), (0,0,0,1,0), and (0,0,0,0,1). Taking the measurement data of sensor A1 as an example, the three-dimensional t-SNE scatter plot corresponding to the input features is shown in the attached figure. Figure 4 As shown in the figure, it can be observed that the principal component data under various working conditions are mixed together, and it is difficult to distinguish the corresponding structural states.
[0061] In the third step, for the signal data from each sensor, a stacked autoencoder (EWOA-DSAE) deep learning model is constructed with the enhanced whale optimization algorithm to optimize the network model parameters. The constructed learning model includes an input layer, an output layer, and three hidden layers with dimensionality reduction configuration. The hyperparameters required to establish the model based on the training data, including the number of neurons in each hidden layer HLN1, HLN2, HLN3, the discarding probability dr, the weight penalty factor λ and the learning rate lr, are optimized and determined by the proposed enhanced whale optimization algorithm. The parameters of the enhanced whale optimization algorithm are set as follows: whale group size Np=30, maximum number of iterations Nmax=100, a0=b=1. For the model parameter optimization of sensor A1, the optimization process of each hyperparameter is as shown in the attached figure. Figures 5 to 10 As shown in Figure 2, it can be seen that the values of the six parameters reach the optimum around the 21st iteration. Fig.11 The convergence of the three methods in the process of hyperparameter optimization is compared. The results show that the proposed enhanced whale optimization is optimal in terms of accuracy and convergence.
[0062] For each sensor, the optimized parameters are used to construct an autoencoder model with three hidden layers, denoted as EWOA-DASAE-i, (i=1,2,…,10). After the data of sensor A1 is extracted through three layers of features, the feature results output by the third hidden layer are shown in the attached figure. Fig.12 As shown in the figure, it can be observed that the features corresponding to the five structural health states are completely dispersed after being analyzed and extracted by the established model. This shows that the proposed model has excellent feature extraction ability and clustering performance when processing vibration signals under different working conditions.
[0063] In order to illustrate the accuracy of the proposed EWOA-DSAE relative to other machine learning methods, a comparative study will be conducted using the measured data of A1. The models used for the comparative study include standard DASE with constant preset hyperparameters, one-dimensional convolutional neural network (1D-CNN), and support vector machine (SVM). The damage diagnosis results show that the overall accuracy of the validation samples predicted by the proposed model is 94.4%, which is higher than 91.2% of standard DASE, 93.6% of 1D-CNN, and 89.6% of SVM.
[0064] Table 1 EWOA-DSAE model recognition results corresponding to each sensor:
[0065]
[0066] Table 1 shows the recognition accuracy of all EWOA-DASAE models corresponding to the 10 sensor data in training, validation, and overall data samples. It is worth noting that most models have satisfactory prediction performance (>90%) for the training samples, except for EWOA-DSAE-2, EWOA-DSAE-4, EWOA-DSAE-6, and EWOA-DSAE-10. This is mainly because these four accelerometers are affected by the deployment location, which reduces the quality of the measurement data and the training accuracy of the developed model.
[0067] In the fourth step, the DS algorithm is used to fuse the outputs of multiple EWOA-DSAE models to obtain the final result. Two sensor layout conditions are considered. The first is the three-sensor layout condition of A1, A2, and A10, and the second is the four-sensor layout condition of A5, A6, A7, and A8. Under each condition, only the data collected by the corresponding sensor is used for analysis. Under the first sensor layout condition, the probability results corresponding to the health status of each structure obtained by the constructed EWOA-DASAE-i (i=1, 2, 10) model are used as separate evidence for constructing basic probability assignments in the DS algorithm for fusion. The confusion matrix of the learning model for the three sensor verification data is shown in the attached figure. Fig.13 The results show that the accuracy of structural damage diagnosis for the validation sample has increased to 96.4%. Compared with 84.8% for only EWOA-DASAE-2 or 76.8% for only EWOA-DASAE-10, the accuracy of the three sensor fusion results has been greatly improved. Similarly, considering the second sensor layout condition, the validation data confusion matrix of the fusion EWOA-DASAE-i (i = 5, 6, 7, 8) model analysis is shown in the attached figure. Fig.14 The results show that the structural damage diagnosis accuracy of the verification sample has increased to 98.8%, which is higher than the recognition accuracy of each sensor in Table 1.
Claims
1. A structural damage diagnosis method based on optimized stacked autoencoders and multi-signal fusion, It is characterized in that The following steps are involved: S1: First, a structural monitoring system is deployed, and n acceleration sensors are used to obtain the vibration response data of the structure in different health states; S2: preprocessing the monitored vibration response data to obtain a principal component data set of frequency domain response; S3: For the principal component data set of the signal of each sensor in S2, a stacked autoencoder EWOA-DSAE deep learning model is constructed with the enhanced whale optimization algorithm EWOA to optimize the network model parameters; the deep learning model EWOA-DSAE-i corresponding to each sensor data is constructed using the optimized hyperparameters, i = 1, 2, ..., n; the probability value of accurate prediction and the preliminary diagnosis results of structural damage are calculated; The enhanced whale optimization algorithm in step S3 performs network model hyperparameter optimization, and the specific process is as follows: 1) An adaptive inertia weight is added to the original whale optimization algorithm to enhance the local search capability of the algorithm. The adaptive inertia weight can be expressed as: Among them, iw max and iw min Represent the maximum and minimum inertia weights respectively; t represents the current iteration number; T max Indicates the maximum number of iterations; 2) The Archimedean spiral curve is used to replace the logarithmic spiral curve in the original method. The mathematical expression of the corresponding curve is: Among them, a 0 ,b is a constant that defines the spiral shape; l represents a random number between 0 and 1; 3) The formation of a new whale bubble network can be expressed as: Among them, X*(t) represents the current best position of the whale, that is, the optimal result; X(t) represents the current position of the whale; a is a variable with a value decreasing from 2 to 0; r represents a random number between 0 and 1; a 0 ,b is a constant that defines the spiral shape; l represents a random number between 0 and 1; p is a random number between 0 and 1; A is a coefficient vector; The calculation process of the enhanced whale optimization method is as follows: S2.1: Set the size of the whale group N p and the maximum number of iterations N max , and initialize the algorithm parameters: a, r, l, a 0 , b and p; S2.2: Calculate the fitness values of Np whales and compare them to obtain the optimal individual with the best fitness, represented by X*; S2.3: Enter the main loop of the algorithm: if |A|<1, update the whale's position according to formula (3); if |A|≥1, update the whale's position according to the prey search behavior; S2.4: Re-evaluate the fitness values of all whales in the group to find the global optimal individual with the corresponding position. If the current optimal fitness is better than the previous one, replace the previous optimal position with the current optimal position; otherwise, keep the optimal individual and position unchanged; S2.5: Check the algorithm stopping criteria: Use the maximum iteration as the termination criteria. If the current number of iterations is greater than the maximum number, the algorithm terminates; otherwise, go to step S2.4 and continue the algorithm iteration; S2.6: Output the optimal solution X*; The optimized stacked autoencoder proposed in step S3 is as follows: 1) The enhanced whale optimization method is used to establish an optimized stacked autoencoder EWOA-DSAE for the input data of each sensor. The optimized autoencoder hyperparameters include the number of neurons in each hidden layer HLN, the weight penalty factor λ, the discard probability dr, and the learning rate lr; 2) The DSAE model loss function is used as the objective function required for hyperparameter optimization, which can be expressed as: in, and k Represents the prediction and true situation of the kth sample; x k represents the kth input; N tr represents the number of training samples; I{·} is the indicator function, when the input is true, its value is 1; otherwise, its value is 0; It represents the probability that the prediction result based on softmax classification belongs to the true label, and its expression is: in, represents the connection weight and bias vector of the i-th neuron in the output layer, corresponding to the i-th structural condition; N cc The total number of structural health conditions; S4: The predicted probability corresponding to the measurement signal of each sensor obtained in S3 is used as the basic probability assignment in the Dempster-Shafer DS fusion algorithm to form the final decision on the structural damage diagnosis analysis.
2. According to claim 1, a structural damage diagnosis method based on optimized stacked autoencoders and multi-signal fusion, It is characterized in that The multi-signal fusion based on the Dempster-Shafer fusion algorithm in step S4 specifically uses the EWOA-DSAE-i, i=1, 2, ..., n model probability results corresponding to the measurement signals of each sensor obtained in step S3 as separate evidence for the basic probability assignment in the Dempster-Shafer algorithm, and fuses the results of multiple sensors to form the final structural damage diagnosis.
